Spatial–Frequency Collaborative Decoupling and Adaptive Gating Network for High-Resolution Remote Sensing Image Semantic Segmentation

Semantic segmentation of high-resolution remote sensing images faces challenges due to intra-class heterogeneity and inter-class spectral similarity. Traditional single-domain methods struggle to balance computational efficiency, high-frequency details, and a global receptive field. To address this, we propose the Spatial-Frequency Collaborative Decoupling and Adaptive Gating Network (DSFANet). The model employs a parallel CNN-Transformer architecture: in the spatial domain, a lightweight large kernel attention (LightLKA) module captures local high-frequency boundaries; in the frequency domain, a 2D-FFT-based global filter establishes long-range topological dependencies with an O(Nlog N) complexity. Additionally, a scene context-aware gating (SCAG) module aggregates global channel-wise statistics to produce two scalar gating weights for fusing the spatial and frequency representations, and a linear-complexity attention fusion block (AFB) bridges the semantic gap between deep and shallow layers. Experiments on four heterogeneous datasets (Vaihingen, Potsdam, LoveDA, and GID) show that DSFANet achieves an optimal balance between accuracy and computational cost. Furthermore, interpretability analysis using Grad-CAM and t-SNE confirms that this dual-domain architecture effectively handles local occlusions, reduces semantic aliasing, and improves inter-class separability. Overall, this work provides an efficient, reliable, and interpretable method for large-scale, fine-grained land use mapping. The code can be obtained here: https://github.com/LijianXie/DSFA-Net.